Voxel51 has released version 3.0 of its data platform, adding a new agent tool and expanding support for multimodal sensor data from robots and vehicles.
Key specifications
- Scale: The Voxel51 Agent includes over 45 Model Context Protocol (MCP) tools, more than 80 operators, and 18 open-source skills.
- Runs on: The platform operates via a hosted web app at app.voxel51.com or through self-hosted deployments in cloud, hybrid, on-prem, and air-gapped environments. The company has not disclosed local GPU requirements.
- Performance: No third-party benchmarks are available. Voxel51 relies on internal survey data for performance claims.
- Free tier: Includes the Voxel51 Agent, 30,000 inference requests, and 30 hours of compute per month.
- Limitations: The free tier is restricted to three datasets of 10,000 samples each and 20 GB of total storage.
New features in Voxel51 3.0
The update extends the platform beyond its original focus on visual data with three main additions.
- Time-series multimodal support: Teams can now view multiple sensor streams on a single timeline. Segment-level embeddings allow the system to locate specific moments within long recordings.
- Unified annotation: Labeling covers 2D, 3D, video, and multimodal data. The system includes automated quality checks, annotator metrics, and human-in-the-loop review capabilities.
- Agentic Labeling: Vision-language models generate labels based on natural-language instructions and provided examples.
Supported data formats include MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data, and audio.
Voxel51 Search mechanics
Voxel51 Search functions as a persistent hybrid index. It indexes datasets and computes embeddings automatically, enabling search functionality immediately upon launch.
Semantic retrieval identifies visually or semantically similar moments even when labels are missing. Lexical and structured queries handle exact terms, timestamps, sensor fields, and metadata. Embeddings and structured fields remain in their native form and are combined during the retrieval process.
Vector retrieval runs on LanceDB. It supports vector, scalar, and full-text indexes while separating storage from compute. Teams decide which data to embed, which model to use, and when to run compute tasks. Incremental updates prevent full rebuilds when new samples arrive or schemas change.
Voxel51 Agent capabilities
The Voxel51 Agent chains search, curation, annotation, and evaluation into repeatable workflows.
Example workflows:
- A robotics team searches for specific behaviors across thousands of episodes, then curates balanced training sets.
- A perception team starts from a model failure and pulls similar long-tail scenarios for labeling.
The agent can write code to build custom plugins. All workflows inherit the organisation’s permissions and governance controls.
Data provenance in Voxel51 3.0
Voxel51 3.0 introduces four provenance layers:
- Deployment-level activity log: Records who performed actions and when across the deployment.
- Dataset-level activity log: Tracks how a single dataset evolved over time.
- Sample-level edit history: Logs label, metadata, and property changes for each individual sample.
- Dataset versioning: Creates immutable snapshots to reproduce training and evaluation runs.
Events flow through fiftyone-mq for job scheduling and Activity Core for event tracking.
Custom model integration
Teams can bring their own models. The platform is model-agnostic across embeddings, auto-labeling, inference, and evaluation. An organisation can start with a foundation model and swap in its own domain-specific model later.
Survey findings on data work
Voxel51 cites its 2026 State of Physical AI survey. The data indicates that 58% of “exceptional” teams spend more than half their project time on data work. Only 21% of struggling teams report the same level of effort. This is vendor-run research, so treat the figures as directional indicators rather than absolute facts.
Comparison with Encord and Foxglove
| Feature | Voxel51 3.0 | Encord | Foxglove |
|---|---|---|---|
| Core focus | Search, curate, annotate, evaluate multimodal data | Annotation, curation and data management | Robotics visualization, observability and data management |
| Free tier | 3 datasets, 10,000 samples each, 20 GB | Not disclosed | 10 GB storage, 3 users, 5 devices |
| Paid entry | Team: 8 seats, 2,800 compute hrs/mo (price not disclosed) | Starter (price not disclosed) | Pro: $20/mo + usage, 1 TB storage |
| Built-in agent | Voxel51 Agent | Data agents (Team plan) | Agent sidebar |
| MCP | 45+ MCP tools | Not disclosed | Foxglove MCP (local MCP on paid plans) |
| Search | Hybrid semantic, lexical and metadata | Multi-modal search, querying, filtering | Semantic Search (beta) plus structured search |
| Annotation | 2D, 3D, video, multimodal, Agentic Labeling | Image, video, audio, documents; 3D and LiDAR add-on | Events on time segments |
| Time-series sensor data | MCAP, synced streams, segment embeddings | Sensor streams (see note) | MCAP, ROS 1, ROS 2 native |
| Provenance | Activity logs, edit history, versioning | Not disclosed | Audit logs (not on Free) |
| Self-hosted | On-prem and air-gapped available | VPC and on-prem (add-on or Enterprise) | Self-hosted data on Enterprise; BYO storage on Pro |
Encord’s sensor-stream support is stated in its Series C release. Prices verified October 17, 2026.
What it means
The release shifts the focus from simple storage to active data management. Teams training models on physical systems can now search and curate data using natural language without writing custom scripts. The inclusion of air-gapped self-hosting options suggests a push towards enterprise security compliance. However, the lack of independent benchmarks means performance claims remain unverified by outside auditors.
